Differential Morphological Profile Neural Networks for Semantic Segmentation

Semantic segmentation of overhead remote sensing imagery supports critical applications in mapping, urban planning, and disaster response, yet state-of-the-art segmentation networks are predominantly designed for ground-perspective imagery and do not directly address remote sensing challenges such as extreme scale variation, foreground–background imbalance, and large image sizes. Rather than proposing new architectures, we take an architecture-agnostic approach by incorporating the differential morphological profile (DMP), a multi-scale shape extraction method based on grayscale morphology, as supplementary input to modern segmentation networks. We evaluate two integration strategies: a Direct-In approach, which adapts the input stem to accept DMP channels in place of or alongside RGB data, and a Hybrid DMP dual-stream architecture in which separate RGB and DMP encoders process each modality independently. Experiments on the iSAID, ISPRS Potsdam, and LoveDA benchmark datasets assess multiple DMP differentials and structuring element shapes. Results show that use of the DMP as direct input into models generally under-perform RGB-only baselines, while the Hybrid DMP approach substantially closes this gap and in some cases surpasses baseline performance, with gains varying across object categories. In the strongest case, a Hybrid DMP SegNeXt-S model achieves a gain of +3.19 mIoU over the RGB-only baseline on the ISPRS Potsdam dataset, and Hybrid DMP models outperform the RGB-only baseline on two of the three benchmark datasets evaluated. These findings suggest that DMP features provide complementary shape information that, when properly integrated, can enhance semantic segmentation performance for overhead remote sensing imagery.